Research On Fault Diagnosis Of Wind Turbine Transmission System Based On Deep Learning Algorithm

In recent years,with the progress of technology and the increasing attention of people to the problem of environmental pollution issues,the installed capacity of wind power generation accounts for an increasing proportion of the total installed capacity of power generation.The working environment of wind turbine is bad,and the wind speed and direction are variable.As the operating load of wind turbine is constantly changing,the frequency of wind turbine faults is very high.Therefore,the fault diagnosis of wind turbine has become a hot topic worldwide.The transmission system is the energy transfer station of wind turbine.Among the components of wind turbine,the transmission system has the most faults.In this paper,a method based on deep learning is used to solve the problem of fault diagnosis of the transmission system of wind turbine.The short-time Fourier transform is used to construct time-frequency diagrams.The convolution neural network is used to learn the fault characteristics from the time-frequency diagram under different faults.The simulation results show that this method is effective in both bearing fault data sets and gearbox fault data sets.The main work of this paper is as follows:Firstly,the basic structure and working principle of wind turbine are introduced.The fault characteristics and causes of bearing and gear are analyzed.On the basis of previous research,a fault diagnosis scheme is designed.Fault samples are composed by fault data of bearing and gear cases.The short-time Fourier transform is used to transform the vibration signal into a two-dimensional time-frequency diagram suitable for the input of the convolution neural network.In this paper,hamming window is selected as the window function in the short-time Fourier transform.The window function width is also set.This method avoids the loss of fault feature information in frequency domain.The fault data set is formed by the time-frequency diagram of bearing and gear data.This data set is divided into training set and test set in proportion.The fault diagnosis model of convolution neural network is established according to the size of time-frequency diagram.The main parameters of the network are set and the influence of hyper parameters on the process and results of network training is studied.The network diagnosis results obtained by two parameter updating algorithms,Stochastic gradient descent(SGD)and Adaptive moment estimation(Adam),are compared.The results show that the Adam optimization algorithm has a better diagnostic effect than the stochastic gradient descent algorithm.Adam algorithm designs an independent adaptive learning rate for different parameters by calculating the first order moment estimation and the second order moment estimation of gradient.It speeds up the convergence speed of the loss function and prevents the loss function from falling into the local optimal point.The results of the example show that Adam algorithm is better than SGD algorithm in reverse updating parameters.This method achieves a diagnosis accuracy of 98.77% on the bearing data set,which is 2.13% higher than SGD optimization algorithm.The accuracy of the gearbox data set reaches 96.69%,which is 3.86% higher than SGD optimization algorithm.Finally,this diagnostic method is compared with other traditional methods and the following conclusions are drawn:(1)The diagnosis accuracy of convolution neural network is higher than that of artificial neural network.The number of layers of convolution neural network is higher than that of artificial neural network.It is good at learning the internal nonlinear mapping relationship of data from large quantities of data.(2)The diagnosis effect of convolution neural network with time-frequency graph as input is better than that of convolution neural network with time-domain characteristics as input.This method can learn the fault character information in time domain and frequency domain at the same time.It overcomes the shortcoming that the time domain index is not strong enough to express fault features.(3)Different from stochastic gradient descent algorithm,Adam optimization algorithm can design an independent learning rate for different parameters.It can also make the network convergence faster and better.